TSINGHUA SCIENCE AND TECHNOLOGY

ISSN  1007-0214  14/19 pp594-601

Volume 12, Number 5, October 2007

A Statistical Method for Analysis of Technical Data of

a Badminton Match Based on 2-D Seriate Images

CHEN Bingqi (чρ୫)**, WANG Zhiqiang (ฆᄝஜ)

Department of Mechanical and Electronical Engineering, China Agriculture University, Beijing 100083, China

Abstract: The use of computer vision technology to collect and analyze statistics during badminton matches
or training sessions can be expected to provide valuable information to help coaches to determine which
tactics should be used by a player in a given game or to improve the player’s tactical training. A method
based on 2-D seriate images by which statistical data of a badminton match can be obtained is presented.
Image capture and analysis were performed synchronously using a multithreading technique. The regions of
movement in the images were detected using a temporal difference method, and the trajectories of the
movement regions were analyzed using seriate images. The shuttlecock trajectory was extracted from all
detected trajectories using various characteristic parameters. The stroke type was determined by comparing
the shuttlecock trajectory data with a set of stroke definition data. The algorithm was tested at a training cen-
ter, and the results were compared with baseline data obtained by expert visual inspection using four video
samples, which included approximately 10 000 frames. The shuttlecock trajectory and stroke type were de-
tected correctly in almost 100% of the analyzed video sequences. The average speed of the automated
analysis was approximately 40 frames/s, indicating that the method can be used for real-time analysis dur-
ing a badminton match. The system is convenient for use by a sports coach.

Key words: machine vision; 2-D measurement; badminton match; trajectory tracking; stroke type statistic

time tracking system for sports broadcast applications

Introduction
At present, plays and goals in sporting matches are
usually judged by the human eye, and recording is per-
formed “manually” by human operators. With the re-
cent improvement of computer hardware and technol-
ogy related to image processing, computer vision is in-
creasingly expected to replace manual operations in
sports matches and training.
Several studies have been performed to analyze the
movement of players and the ball in sporting events,
and many developments have been made in the past
few years. Pingali et al. developed a multi-camera real-
Received: 2006-03-09; revised: 2006-12-04
γγTo whom correspondence should be addressed.
E-mail: cbq93@sohu.com; Tel: 86-10-62966687
and proposed techniques for the game of tennis for
tracking the ball and players on video images obtained
from stationary cameras[1,2]. Yu et al. proposed a
framework to determine the location of a ball by look-
ing for and analyzing the ball trajectory from broadcast
soccer match video data[3-6]. Ekin et al. provided a
framework for automatically analyzing and obtaining
statistics of a soccer match in real-time[7]. Utsumi et al.
proposed a detection and tracking method in order to
detect and track the objects necessary to describe a
soccer match[8]. Perš and Kovacic developed a multi-
ple-camera-based player tracking system that could be
used for simultaneously tracking several players in a
large section of a handball court using a series of
384×288 pixels images acquired by fixed cameras[9,10].
Figueroa et al. considered the problem of tracking
 CHEN Bingqi (чρ୫) et alġA Statistical Method for Analysis of Technical Data of …
players during a soccer game through the use of multi-
ple cameras[11]. Ye et al. have proposed a method for
athlete identification by combining the segmentation,
tracking, and recognition procedures into a coarse-to-
fine scheme for jersey number (digital characters on
the sport shirt) detection[12]. Wang et al. proposed an
automatic technique for extracting color models of the
playing surface and the team uniforms that can be used
in higher-level processes such as tracking and recogni-
tion[13]. Qi et al. presented a framework for tracking
sports players in videos recorded by static domestic
cameras[14]. Zaveri et al. proposed an algorithm for de-
tection and tracking of small and fast moving objects,
such as a pingpong or a cricket ball, in sports video se-
quences[15].
In badminton matches and in training, in order to se-
lect the tactics required to deal with different oppo-
nents, the coach must have a good statistical know-
ledge of the tactics of the player. The ultimate goal of
this study was to develop a system capable of real-time
statistical analysis of data collected either during train-
ing or match-play, to provide information to assist
training and to select match tactics.
The use of 3-D information should be able to pro-
vide accurate statistical data for analysis of a badmin-
ton match. However, obtaining 3-D information re-
quires the use of more than two cameras, and real-time
analysis requires high-capacity processors, the size and
volume of which are generally too large for a practical
implementation. Therefore, 3-D detection is not suit-
able for badminton coaches, as the equipment must be
readily, transportable to remote match locations. In the
present study, therefore, we develop a 2-D measure-
ment and analysis system, which requires only one
camera and a laptop personal computer.
1 Method
The system can be broken into a number of steps as
follows.
(1) Capture of 2-D images of the badminton court
that are suitable for use in measuring the shuttlecock
trajectory using one camera.
(2) Extraction of moving objects from 2-D seriate
images.
(3) Classification of moving objects and connection
of their tracks.
(4) Extraction of the shuttlecock trajectory from
entire tracks.
595
(5) Calculation of the statistics for the stroke type by
comparing the characteristics of the detected shuttle-
cock trajectory and the definition of the stroke type.
1.1 Equipment
A Basler A601f series progressive scan CMOS (com-
plementary metal-oxide-semiconductor transistor)
camera was used to acquire sample images. The cam-
era was designed for industrial applications. The pri-
mary features of the camera are as follows. The camera
matches the 1394 TA (transmission adapter) digital
camera specification. The maximal frame rate at full
resolution is 60 frames/s. The asynchronous full frame
is shuttered electronically. The camera has square pix-
els, and the frame size is 659×493 pixels. The camera
was used to capture monochrome images.
Capturing and processing of each image were per-
formed using a notebook personal computer with a
Pentium 2.4-GHz CPU. The algorithm was developed
based on the platform of a 2-D moving image meas-
urement system, which was provided by Beijing Mod-
ern Fubo Technology Ltd., implemented using Micro-
soft Visual C++ 6.0.
1.2 Sampling a moving shuttlecock
1.2.1 Preparation for the sampling program
The application for synchronous capture, analysis, and
data storage uses Windows multithreading technology.
When the application is started, the captured thread is
started and captured images are displayed on the
screen. Then, when analysis and/or saving has begun,
thread analysis and/or saving is started, and image
analysis and/or saving is carried out synchronously.
1.2.2 Establishing the camera and demarcating
the badminton court
In the present study we attempt to develop a real-time
measurement algorithm for a moving shuttlecock in an
actual match. Therefore, the image samples used in the
present study should simulate scenes of an actual
match as closely as possible. In order to capture the en-
tire scene of the badminton court, the camera was
mounted 5 m from the corner of the court at a height of
4 m from the ground and at an angle of approximately
45q from the vertical.
Before the image analysis, eight characteristic coor-
dinates of the badminton court were selected
manually, as shown in Fig. 1. The representational
 596
points comprise the two top ends of the badminton net
(point 1,3), the two feet of the net frame (point 2,4),
and the four corners of the court near the net (point 5-8).
Fig. 1 Representational points of the net and the field
1.3 Trajectory tracking and analysis
1.3.1 Concept of direction number
In the present study we introduce the concept of a di-
rection number, which is used to determine the flight
direction and stroke type. The direction number is de-
fined as follows. When the x coordinate of a point on
the trajectory is increased relative to previous point,
the direction number is positive. When the x coordinate
of a point on the trajectory is decreased relative to pre-
vious point, the direction number is negative. The ap-
plication states that the initial direction number is zero.
When tracking is complete, the application can deter-
mine the direction and length of the trajectory accord-
ing to the value of the direction number.
1.3.2 Object detection
Detection of differences between current images and
the background is the most common method used for
detection of the moving area in image processing.
However, this method requires a background image in
which there are no moving objects, which is difficult to
obtain in a match environment. In the present study we
therefore detect the moving area using the difference
between two consecutive frames in the seriate data.
This detection method is hereinafter referred to as the
temporal difference method. The temporal difference
method includes the following procedures. First, the
original image is defined as
X t {x(ti , j ) 0 İ i  M ;0 İ j  N } (1)
where M is the width of the image and N is the height
of the image.
Second, the temporal difference image is defined as
­ 255, X t  X t 1 ! n;
Rt ®
¯0, X t  X t 1  n (2)
where n is a threshold that can be modified and R t is
a binary image.
Tsinghua Science and Technology, October 2007, 12(5): 594-601
The outline of each white pixel block is measured on
the binary image. According to the coordinates of the
outline, the barycenter of each white pixel block is cal-
culated as
1 n 1 1 n 1
X ¦
ni 0
xi , Y ¦ yi
ni 0
(3)
where X is the x coordinate of the barycenter, Y is the y
coordinate of the barycenter, xi and yi are the x and y
coordinates of point i on the outline, and n is the num-
ber of points on the outline.
1.3.3 Trajectory detection
(1) Establishing the trajectory record
The trajectory record for each frame includes the
following information:
(a) The coordinates of the trajectory in the current
scene;
(b) The angle between the x-axis and the line con-
necting the current point with the previous point on the
trajectory;
(c) The distance between the current point and the
previous point on the trajectory;
(d) The current direction number.
(2) Trajectory matching
Cmi ( x, y ) is used to describe the barycenter coordi-
nate of area i in frame m. The distance L and the angle
A from C mi ( x, y ) to Cmmax
1 ( x , y ) , which is a point be-
longing to the longest trajectory in the m1 frame, are
calculated. If the distance L is between the established
longest and shortest distances, and the angle A is be-
tween the established largest and smallest angles, then
Cmi ( x, y ) is judged to be a point on the longest trajec-
tory. The point Cmi ( x, y ) is then recorded in the long-
est trajectory record and the direction number of the
longest trajectory is modified. If Cmi ( x, y ) is judged
not to belong to the longest trajectory, and if Cmi ( x, y )
is matched another trajectory, it is recorded as belong-
ing to the matching trajectory record. If Cmi ( x, y )
does not match any trajectory on the m1 frame, then a
new trajectory record starting at Cmi ( x, y ) is created.
In the trajectory matching, several points on the
same frame may appear to match the same trajectory.
In this case, the point nearest to the trajectory is re-
corded as belonging to the trajectory.
If none of the barycenters on frame m matches any
of the barycenters on frame m1, then the longest
 CHEN Bingqi (чρ୫) et alġA Statistical Method for Analysis of Technical Data of … 597

trajectory up to this point is analyzed in order to de- position of the trajectory start point, the position of the
termine whether the longest trajectory is the shuttle- trajectory end point, the direction and the distance of
cock trajectory. the line from the trajectory start point to the end point,
the trajectory camber, the average distance between
1.4 Detecting the shuttlecock track
points on the trajectory, the location of the highest
Tmax denotes the longest trajectory among all of the point on the trajectory, the distance between the trajec-
tory start point and the end point to the players, the dis-
trajectories, and O1 , O2 ,..., Om denote all of the proper-
tances from the start point and the end point to the
ties that are used to evaluate whether a trajectory is the
badminton net, the direction number of the trajectory,
shuttlecock trajectory. These properties include the tra-
the coordinates of the badminton net, the coordinates
jectory length, the trajectory camber, and the trajectory
of the court contour, and the locations of the trajectory
direction number. The reliability parameter \ (Tmax )
start and end points in the badminton court.
of the shuttlecock trajectory is defined as Flow chart of the detection system is shown in Fig. 2.
m
\ (Tmax ) ¦\
i
i (Oi ) (4)

If the reliability is greater than some established

threshold, the trajectory is judged to be the shuttlecock
trajectory. The shuttlecock trajectory is then analyzed
to determine the stroke type. If the stroke type is de-
termined, all of the trajectory records are deleted, and
the detection process is restarted. If the longest trajec-
tory is not the shuttlecock trajectory, all of the trajec-
tory record is stored in memory and the detection proc-
ess is restarted.
During the detection process the shuttlecock may
temporarily leave and then reenter the visual field of
the camera. Therefore, whether or not a detected shut-
tlecock trajectory is part of the previous shuttlecock
trajectory must be determined. If the distance from the
end of the previous shuttlecock trajectory to the start of
the current shuttlecock trajectory is within a given
range, then the current shuttlecock trajectory is deter-
mined to be part of the previous trajectory; otherwise,
the current shuttlecock trajectory is classified as being
unrelated to the previous trajectory.

1.5 Judging the stroke type of a moving

shuttlecock

After detection of the shuttlecock trajectory, the stroke

type is determined by comparing the data of the de-
tected shuttlecock trajectory and the stroke definition.
In the present study, the defined stroke types include
lob, raise, smash, backstop, punch, and whip.
A number of parameters are used to determine the
shuttlecock type. These are the trajectory length, the Fig. 2 Flow chart of the detection system
 598
2 Results and Discussion
2.1 Sampling the moving shuttlecock
The real-time statistical data collection system requires
the captured image to cover the entire badminton court,
and sufficient space between the top of the net to the
top of the image is required in order to achieve re-
quired measurement precision. As shown in Fig. 3 for
the captured images, the top of the net was located just
above the center of the image and the field of vision
covered approximately the entire badminton court, in-
dicating that the requirements for camera position and
angle were satisfied.
Fig. 3 Badminton field
Tests revealed that a capture frequency of 40
frames/s was adequate for capturing even the fastest
shuttlecock during a badminton match. Therefore, in
the present study, the capture frequency was set to 40
frames/s. The algorithm was analyzed using four video
samples and was tested at a training center. The frame
counts for the four video samples were 394, 1204, 3108
and 6020. The average saving speed for the video sam-
ples was 38 frames/s; this can be improved by using a
higher-capacity computer.
2.2 Measuring the moving shuttlecock
In obtaining the position information of the badminton
court and net, a small number of characteristic points
have to be demarcated before measuring the shuttle-
cock trajectory. This position information is important
for determining the stroke type and for delimiting the
badminton court in order to avoid noise from outside
the badminton court. Figure 3 shows eight characteris-
tic points that were obtained by manually clicking on
the image. The lines indicate the area of the badminton
Tsinghua Science and Technology, October 2007, 12(5): 594-601
court calculated using these characteristic points.
Statistical data for the athletes on both sides of the
net must be recorded while analyzing the shuttlecock
trajectory. The direction of the shuttlecock trajectory
can be determined according to the direction number at
the end of the trajectory. If the direction number is
positive, indicating that the shuttlecock moved from
left to right, then the measurement data should be
added to the record of the athlete on the left side of the
net. Otherwise, the measurement data should be added
to the record of the athlete on the right side of the net.
Tests revealed that use of a direction number to deter-
mine the trajectory direction provided correct results.
The direction number was not only used to determine
the trajectory direction, but also the length of the tra-
jectory.
Figure 4 shows a difference binary image together
with the two original images used to create it. Original
image 2 appeared after the Original image 1 in the im-
age sequence. The threshold for the binary realization
was set at 5 for the difference image. Moving parts in
the images, such as the shuttlecock and players, are
denoted as white pixels; the static parts, such as the
badminton court and the badminton net, are denoted as
a black background in the binary image. Such binary
images were obtained continuously from the image se-
quence, and each barycenter of each moving object
was determined.
The large “+” symbols on the net line in the binary
image (Fig. 4c) indicate the center of the net line. The
two large “+” symbols on the players indicate their
barycenters. The smaller “+” symbols on each of the
white pixels indicate the barycenters of these pixels.
The barycenter of each player was calculated correctly
by using the block barycenter of each white pixel in
the fixed measurement area. The borderlines of the
measurement area, which are denoted by the white
lines on the binary image (Fig. 4c), are drawn on the
image view.
Figure 5 shows the barycenters for a single image
and also the superimposed barycenters for a series of
images. It is difficult to separate the barycenter of the
shuttlecock in one single image from those of other
moving features (Fig. 5a), but the trajectory of the shut-
tlecock barycenter can be clearly identified from inspec-
tion of the superimposed image series (Fig. 5b). The tra-
jectory of other moving features, such as the players and
 CHEN Bingqi (чρ୫) et alġA Statistical Method for Analysis of Technical Data of … 599

Fig. 5 Barycenters of objects

Fig. 4 Original images and binary image

their rackets, differ from the shuttlecock trajectory (Fig.

5b). Diagnostic data were recorded by connecting each
barycenter with a trajectory, and the shuttlecock trajec-
tory was determined by the analysis of these diagnostic
data.
Figure 6 shows two samples of the moving shuttle-
cock trajectory detection. Figure 6a shows an intact
shuttlecock trajectory, and Fig. 6b shows an interrupted
shuttlecock trajectory. The trajectory data are subse-
quently used to judge the trajectory type of the moving
shuttlecock.
Fig. 6 Samples of detected ball track
 600 Tsinghua Science and Technology, October 2007, 12(5): 594-601

The detected trajectory data were compared with a video samples, which included a total of approximately
set of stroke type definitions to determine the stroke 10 000 scenes. The results are compared in Table 1.
type. The results obtained by the algorithm were com- Video samples 1, 2, 3, and 4 had 394, 1204, 3108, and
pared with those obtained from visual inspection of the 6020 scenes, respectively.
Table 1 Measurement results by the algorithm (A) and visual inspection (V)
Lob Raise Smash Backstop Punch Whip
Video sample
A V A V A V A V A V A V
1   2 2 2 2   1 1 1 1
2 2 2 9 9 3 3 1 1 6 6 1 1
3 9 9 9 9 2 2 3 3 8 8 20 20
4 6 6 28 29 16 16 6 6 13 13 4 4

As shown in Table 1, the detection rate of the shut-
tlecock track was close to 100%. In only one case, a
raise stroke, the shuttlecock track was not recorded. In
this case, the shuttlecock trajectory was in fact deter-
mined but ignored because the shuttlecock trajectory
was not the longest trajectory among the group of tra-
jectories.
The test results indicate that the algorithm can be
used to obtain statistical data of player tactics during a
badminton match.
3 Conclusions
This paper described a method to obtain statistical data
of a badminton match based on 2-D seriate images.
Moving portions of the images, including the shuttle-
cock, rackets and players, are detected and labeled as
white pixel blocks based on an image difference
method. The outline of each white pixel block is meas-
ured in the binary image, and the barycenter is calcu-
lated from this outline data. Each barycenter point is
connected to a moving track according to several data.
The shuttlecock trajectory is extracted from the trajec-
tory data. After the shuttlecock trajectory is determined,
the stroke type is determined by comparing the trajec-
tory data to the definition data for various stroke types.
This process is continued until the end of each badmin-
ton match. The output is a set of statistical data of the
stroke type for each player.
The algorithm can be used not only for the training
and sports match, but also for other automated analysis,
such as the track of the insects. In the future, we hope
to accomplish three objectives: (1) improvement of the
algorithm to identify the objective track even when the
track is greatly disturbed; (2) improvement of the algo-
rithm to allow a more detailed definition of the stroke
type; and (3) improvement of the algorithm to enable
the score of the badminton match to be judged.
Acknowledgements
This study supported by Beijing Modern Fubo Technology Ltd.
We thank Dr. Sun Ming, at China Agriculture University, Ms.
Wang Kaizhen and Mr. Zhou Zhihui, at Beijing Sports Univer-
sity, and Mr. Chen Yao, at the Pingpong and Badminton Center
of China National Sports Ministry, for their help in this study.
References
[1] Pingali G, Opalach A, Jean Y. Ball tracking and virtual re-
plays for innovative tennis broadcasts. In: Proceedings
15th International Conference on Pattern Recognition.
Washington, DC, USA: IEEE Compter Society Press,
2000: 152-156.
[2] Pingali G, Jean Y, Carlbom I. Real-time tracking for en-
hanced sports broadcasts. In: Proceedings of CVPR98.
Washington, DC, USA: IEEE Compter Society Press,
1998: 260-265.
[3] Yu Xinguo, Xu Changsheng, Leong Hon Wai, et al. Tra-
jectory-based ball detection and tracking with applications
to semantic analysis of broadcast soccer video. In: Pro-
ceedings of the 11th ACM International Conference on
Multimedia. New York, NY, USA: ACM Press, 2003:
11-20.
[4] Yu X, Tian Q, Wan K W. A novel ball detection frame-
work for real soccer video. In: Proceedings of ICME.
Washington, DC, USA: IEEE Computer Society Press,
2003: 265-268.
[5] Yu X, Xu C, Tian Q, et al. A ball tracking framework for
 CHEN Bingqi (чρ୫) et alġA Statistical Method for Analysis of Technical Data of …
broadcast soccer video. In: Proceedings of ICME. Wash-
ington, DC, USA: IEEE Computer Society Press, 2003:
273-276.
[6] Yu Xinguo, Sim Chern-Horng, Wang Jenny Ran, et al. A
trajectory-based ball detection and tracking algorithm in
broadcast tennis video. In: Proceedings of ICIP. New York,
NY, USA: ACM Press, 2004: 1049-1052.
[7] Ekin A, Tekalp A M, Mehrotra R. Automatic soccer video
analysis and summarization. IEEE Transactions on Image
Processing, 2003, 7(12): 796-807.
[8] Utsumiy O, Miurayy K, Idez I, et al. An object detection
method for describing soccer games from video. In: IEEE
International Conference on Image Processing. 2002:
45-48.
[9] Perš J, Kovacic S. Computer vision system for tracking
players in sports games. In: Proceedings of Int’l Work-
Shop Image Signal Processing Analysis. Pula, Croa-tia,
2000: 177-182.
[10] Perš J, Kovacic S. Tracking people in sport: Making use of
partially controlled environment. In: Proceedings of the
9th International Conference on Computer Analysis of Im-
ages and Patterns. London, UK: Springer-Verlag, 2001:
374-382.
601
[11] Figueroa P, Leite N, Barros R M L, et al. Tracking soccer
players aiming their kinematical motion analysis. Com-
puter Vision and Image Understanding, 2006, 101(2):
122-135.
[12] Ye Qixiang, Huang Qingming, Jiang Shuqiang, et al. Jer-
sey number detection in sports video for athlete identifica-
tion. In: Proceedings of SPIEüThe International Society
for Optical Engineering, Visual Communications and Im-
age Processing. 2005: 1599-1606.
[13] Wang Lei, Zeng Boyi, Lin S, et al. Automatic extraction of
semantic colors in sports video. In: IEEE International
Conference on Acoustics, Speech and Signal Processing.
Canada, 2004: 617-620.
[14] Qi Fei, Luo Yupin, Hu Dongcheng. Visual tracking of
players through occlusions in low resolution. In: 6th
IASTED International Conference on Signal and Image
Processing. Singapore, 2004: 375-380.
[15] Zaveri M A, Merchant S N, Desai U B. Small and fast
moving object detection and tracking in sports video se-
quences. In: IEEE International Conference on Multimedia
and Expo. Taipei, China, 2004: 1539-1542.